1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
15 #include "bluesky-private.h"
17 /* Core filesystem. Different proxies, such as the NFSv3 one, interface to
18 * this, but the core actually tracks the data which is stored. So far we just
19 * implement an in-memory filesystem, but eventually this will be state which
20 * is persisted to the cloud. */
22 /* Return the current time, in microseconds since the epoch. */
23 int64_t bluesky_get_current_time()
26 g_get_current_time(&t);
27 return (int64_t)t.tv_sec * 1000000 + t.tv_usec;
30 /* Update an inode to indicate that a modification was made. This increases
31 * the change counter, updates the ctime to the current time, and optionally
32 * updates the mtime. This also makes the inode contents subject to writeback
33 * to storage in the future. inode must already be locked. */
34 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
36 int64_t now = bluesky_get_current_time();
37 inode->change_count++;
42 if (inode->change_time == 0)
43 inode->change_time = now;
46 if (bluesky_options.writethrough_cache)
47 bluesky_file_flush(inode, NULL);
50 g_mutex_lock(inode->fs->lock);
51 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
52 inode->unlogged_list = bluesky_list_prepend(&inode->fs->unlogged_list, inode);
53 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
54 inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
55 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
56 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
57 g_mutex_unlock(inode->fs->lock);
59 bluesky_cloudlog_unref(inode->committed_item);
60 inode->committed_item = NULL;
63 /* Unfortunately a glib hash table is only guaranteed to be able to store
64 * 32-bit keys if we use the key directly. If we want 64-bit inode numbers,
65 * we'll have to allocate memory to store the 64-bit inumber, and use a pointer
66 * to it. Rather than allocate the memory for the key, we'll just include a
67 * pointer to the 64-bit inum stored in the inode itself, so that we don't need
68 * to do any more memory management. */
69 static guint bluesky_fs_key_hash_func(gconstpointer key)
71 uint64_t inum = *(const uint64_t *)key;
75 static gboolean bluesky_fs_key_equal_func(gconstpointer a, gconstpointer b)
77 uint64_t i1 = *(const uint64_t *)a;
78 uint64_t i2 = *(const uint64_t *)b;
82 /* Filesystem-level operations. A filesystem is like a directory tree that we
83 * are willing to export. */
84 BlueSkyFS *bluesky_new_fs(gchar *name)
86 BlueSkyFS *fs = g_new0(BlueSkyFS, 1);
87 fs->lock = g_mutex_new();
88 fs->name = g_strdup(name);
89 fs->inodes = g_hash_table_new(bluesky_fs_key_hash_func,
90 bluesky_fs_key_equal_func);
91 fs->next_inum = BLUESKY_ROOT_INUM + 1;
92 fs->store = bluesky_store_new("file");
93 fs->flushd_lock = g_mutex_new();
94 fs->flushd_cond = g_cond_new();
95 fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
96 bluesky_cloudlog_equal);
98 fs->log_state = g_new0(BlueSkyCloudLogState, 1);
99 fs->log_state->data = g_string_new("");
104 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
106 BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
108 BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
111 fs->log = bluesky_log_new("journal");
113 g_print("Loaded filesystem superblock\n");
115 fs->name = g_strdup(name);
118 bluesky_string_unref(data);
121 g_print("Initializing fresh filesystem\n");
122 BlueSkyFS *fs = bluesky_new_fs(name);
124 fs->log = bluesky_log_new("journal");
127 BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
131 bluesky_insert_inode(fs, root);
132 bluesky_inode_update_ctime(root, TRUE);
134 bluesky_inode_do_sync(root);
135 bluesky_superblock_flush(fs);
140 /* Inode reference counting. */
141 void bluesky_inode_ref(BlueSkyInode *inode)
143 g_atomic_int_inc(&inode->refcount);
146 void bluesky_inode_unref(BlueSkyInode *inode)
148 if (g_atomic_int_dec_and_test(&inode->refcount)) {
149 if (bluesky_verbose) {
150 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
151 "Reference count for inode %"PRIu64" dropped to zero.",
155 /* Sanity check: Is the inode clean? */
156 if (inode->change_commit < inode->change_count
157 || inode->accessed_list != NULL
158 || inode->unlogged_list != NULL
159 || inode->dirty_list != NULL) {
160 g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64"; accessed_list = %p; dirty_list = %p)\n", inode->change_commit, inode->change_count, inode->accessed_list, inode->dirty_list);
163 /* These shouldn't be needed, but in case the above warning fires and
164 * we delete the inode anyway, we ought to be sure the inode is not on
166 g_mutex_lock(inode->fs->lock);
167 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
168 bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
169 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
170 g_mutex_unlock(inode->fs->lock);
172 /* Free file type specific data. It should be an error for there to be
173 * dirty data to commit when the reference count has reaches zero. */
174 switch (inode->type) {
175 case BLUESKY_REGULAR:
176 for (int i = 0; i < inode->blocks->len; i++) {
177 BlueSkyBlock *b = &g_array_index(inode->blocks,
179 if (b->type == BLUESKY_BLOCK_DIRTY) {
180 g_error("Deleting an inode with dirty file data!");
182 bluesky_cloudlog_unref(b->ref);
183 bluesky_string_unref(b->dirty);
185 g_array_unref(inode->blocks);
188 case BLUESKY_DIRECTORY:
189 g_hash_table_destroy(inode->dirhash);
190 g_hash_table_destroy(inode->dirhash_folded);
191 g_sequence_free(inode->dirents);
194 case BLUESKY_SYMLINK:
195 g_free(inode->symlink_contents);
202 g_mutex_free(inode->lock);
208 /* Allocate a fresh inode number which has not been used before within a
209 * filesystem. fs must already be locked. */
210 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
214 inum = fs->next_inum;
217 bluesky_superblock_flush(fs);
222 /* Perform type-specification initialization of an inode. Normally performed
223 * in bluesky_new_inode, but can be separated if an inode is created first,
224 * then deserialized. */
225 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
230 case BLUESKY_REGULAR:
231 i->blocks = g_array_new(FALSE, TRUE, sizeof(BlueSkyBlock));
233 case BLUESKY_DIRECTORY:
234 i->dirents = g_sequence_new(bluesky_dirent_destroy);
235 i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
236 i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
243 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
244 BlueSkyFileType type)
246 BlueSkyInode *i = g_new0(BlueSkyInode, 1);
248 i->lock = g_mutex_new();
253 bluesky_init_inode(i, type);
258 /* Retrieve an inode from the filesystem. Eventually this will be a cache and
259 * so we might need to go fetch the inode from elsewhere; for now all
260 * filesystem state is stored here. inode is returned with a reference held
262 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
264 BlueSkyInode *inode = NULL;
270 g_mutex_lock(fs->lock);
271 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
274 bluesky_inode_fetch(fs, inum);
275 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
279 bluesky_inode_ref(inode);
281 /* FIXME: We assume we can atomically update the in-memory access time
283 inode->access_time = bluesky_get_current_time();
286 g_mutex_unlock(fs->lock);
291 /* Insert an inode into the filesystem inode cache. fs should be locked. */
292 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
294 g_hash_table_insert(fs->inodes, &inode->inum, inode);
297 /* Start writeback of an inode and all associated data. */
298 void bluesky_inode_start_sync(BlueSkyInode *inode)
300 GList *log_items = NULL;
302 if (inode->type == BLUESKY_REGULAR)
303 bluesky_file_flush(inode, &log_items);
305 BlueSkyCloudLog *cloudlog = bluesky_serialize_inode(inode);
307 bluesky_cloudlog_unref(inode->committed_item);
308 inode->committed_item = cloudlog;
310 bluesky_cloudlog_sync(cloudlog);
311 bluesky_cloudlog_ref(cloudlog);
312 log_items = g_list_prepend(log_items, cloudlog);
314 /* Wait for all log items to be committed to disk. */
315 bluesky_log_finish_all(log_items);
317 /* Mark the inode as clean */
318 inode->change_commit = inode->change_count;
319 inode->change_time = 0;
320 g_mutex_lock(inode->fs->lock);
321 bluesky_list_unlink(&inode->fs->unlogged_list, inode->unlogged_list);
322 inode->unlogged_list = NULL;
323 g_mutex_unlock(inode->fs->lock);
326 /* Write back an inode and all associated data and wait for completion. Inode
327 * should already be locked. */
328 void bluesky_inode_do_sync(BlueSkyInode *inode)
330 if (bluesky_verbose) {
331 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
332 "Synchronous writeback for inode %"PRIu64"...", inode->inum);
334 bluesky_inode_start_sync(inode);
335 if (bluesky_verbose) {
336 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
337 "Writeback for inode %"PRIu64" complete", inode->inum);
341 static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
343 if (bluesky_verbose) {
344 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
345 "Completing fetch of inode %"PRIu64"...", inode->inum);
348 if (async->result != 0
349 || !bluesky_deserialize_inode(inode, async->data->data))
351 if (bluesky_verbose) {
352 g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
353 " failed to load inode, cleaning up");
355 g_mutex_lock(inode->fs->lock);
356 g_hash_table_remove(inode->fs->inodes, &inode->inum);
357 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
358 inode->accessed_list = NULL;
359 g_mutex_unlock(inode->fs->lock);
360 bluesky_inode_unref(inode);
363 inode->access_time = bluesky_get_current_time();
364 g_mutex_lock(inode->fs->lock);
365 bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
366 inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
367 g_mutex_unlock(inode->fs->lock);
369 g_mutex_unlock(inode->lock);
370 bluesky_inode_unref(inode);
373 /* Fetch an inode from stable storage. The fetch can be performed
374 * asynchronously: the in-memory inode is allocated, but not filled with data
375 * immediately. It is kept locked until it has been filled in, so any users
376 * should try to acquire the lock on the inode before accessing any data. The
377 * fs lock must be held. */
378 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
381 sprintf(key, "inode-%016"PRIx64, inum);
383 BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
384 inode->change_count = 0;
385 bluesky_inode_ref(inode); // Extra ref held by fetching process
386 g_mutex_lock(inode->lock);
387 bluesky_insert_inode(fs, inode);
389 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
390 async->op = STORE_OP_GET;
391 async->key = g_strdup(key);
393 bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
394 bluesky_store_async_submit(async);
396 if (bluesky_options.sync_inode_fetches) {
397 bluesky_store_async_wait(async);
400 bluesky_store_async_unref(async);
403 /* Synchronize filesystem superblock to stable storage. */
404 void bluesky_superblock_flush(BlueSkyFS *fs)
407 GString *buf = g_string_new("");
408 bluesky_serialize_superblock(buf, fs);
409 BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
411 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
412 async->op = STORE_OP_PUT;
413 async->key = g_strdup("superblock");
415 bluesky_store_async_submit(async);
416 bluesky_store_async_unref(async);
418 //bluesky_store_sync(fs->store);